SillyTavern-Aether

A SillyTavern extension that lets ST drive an AetherRoom chat backend. It rewrites outgoing text-completion prompts into the AetherRoom format, and exposes per-chat style controls.

Installation

1. Install the extension

In SillyTavern: Extensions → Install extension → paste this repo's URL. Reload the page after the install finishes.

2. Install the companion tokenizer plugin (required)

The extension will not generate responses without the GLM tokenizer running server-side. Accurate token counts are how it keeps the context formatte right.

Enable server plugins in config.yaml:

enableServerPlugins: true

Restart SillyTavern, then from the SillyTavern checkout directory:

node plugins.js install https://github.com/AetherDreams/SillyTavern-Aether-Tokenizer

Restart again. The extension probes the plugin on load and shows a setup banner at the top of the chat pane if it can't reach it. Once the plugin is healthy, the banner is replaced by the AetherRoom top bar.

Setup

1. Get a NovelAI API token. Follow NovelAI's account-settings guide to generate one. Keep it handy for step 2.

2. Configure the SillyTavern API connection. Open the API Connections panel and set:

   Field	Value
   API	Text Completion
   API type	Generic (OpenAI-compatible)
   API URL	https://text.novelai.net/oa
   API token	the token from step 1
   Model name	xialong-v1

   Aether uses text completion because the model's chat template has whitespace details only this extension knows how to produce — chat completion endpoints would mangle them.

3. Pick your character as usual.

4. Open a chat. The AetherRoom top bar appears at the top of the chat pane with selectors for Style, Relationship, Response length, and Japanese. These are saved per-chat and survive reload.

The prompt rewrite only runs when all of the following are true:

• ST's API is set to Text Completion.
• The API URL contains novelai.net.
• The model name contains xialong.

Switching to any other API, URL, or model disables the rewrite automatically — so you can keep the extension installed alongside other text-completion backends without it interfering. The auto-detect patterns are in src/constants.js if you need to tweak them.

What it does to your prompts

• Folds the character's description and personality fields into one Personality block.
• Puts the character's scenario field into its own scenario block so the model treats it as situational context, not character traits.
• Parses mes_example (<START>-delimited) into structured example quotes.
• Sends World Info entries as named context blocks: constant entries pinned at the top, dynamic ones bundled near the recent end of history to optimize performance.
• Supports NovelAI-like context rollover.

A note on the mapping

ST character cards and AetherRoom contacts don't have a clean one-to-one relationship — ST treats character authorship as free-form prose, while AetherRoom contacts have structured fields (Gender, Pronouns, Species, Appearance, Personality, Relationship, etc.). The extension does the best it can with what's on the card:

• description and personality are concatenated into Personality. Anything the card author wanted to call out as appearance, pronouns, or species stays inside that single blob rather than getting routed to a dedicated AetherRoom field.
• scenario becomes the Scene field of the scenario block.
• Card metadata that the v2 spec forbids prompts to use (tags, creator_notes, creator) is intentionally not sent.

The result reads correctly to the model but won't match a hand-authored AetherRoom contact card field-for-field. Expect some loss of structure on cards designed around the AetherRoom schema; ST-authored cards work fine.

Emotion display

Model responses come back as multi-bubble structured text with an emotion per bubble. The extension parses each rendered character message and:

• Splits it into one visual bubble per sub-message.
• Strips the leading {character}: name prefix and the four-space continuation indents so each bubble reads as clean prose.
• Renders the bubble's emotion as a large, semi-transparent emoji in the upper-right of the bubble background. Hovering the emoji shows the emotion name.

The stored chat data is kept in the raw structured form. When you click Edit on a message you'll see the raw name: / Emotion: lines and can fix any wrong emotions by hand — saving re-renders the bubbles.

Recommended sampler preset

A presets/AetherRoom.json text-completion preset ships with this repo. It's the recommended sampler config for the AetherRoom backend:

Active	Disabled
temperature = 0.85	rep_pen = 1 (no penalty)
top_p = 0.95	presence_pen = 0, freq_pen = 0
top_k = 250	min_p, top_a, tfs, typical_p, mirostat, dry, xtc, dynatemp, … all off

To install it:

1. Copy presets/AetherRoom.json from the installed extension folder into your SillyTavern user data folder at data/<user>/TextGen Settings/.
2. Reload SillyTavern.
3. In the Text Completion Preset dropdown, select AetherRoom.

You can still use any other preset you like — this one's just our default recommendation. Two things the extension always enforces regardless of preset, because they're correctness rather than taste:

• max_tokens and max_new_tokens are forced to 1536 (512 × 3).
• truncation_length is bumped to 36864 (= base context + rollover window) so backends that honour it don't crop our prompt down to ST's default 8192-token context.

Troubleshooting

• Setup banner stuck on "tokenizer plugin not detected". Verify enableServerPlugins: true is in config.yaml, the plugin is in ./plugins/, and SillyTavern has been restarted since installing it. Click "Retry" on the banner.
• Generation refuses to start. Same root cause as above — the extension blocks generation when the tokenizer is unreachable rather than sending the wrong format to your backend.
• Top bar doesn't appear. Make sure a chat is open and the extension is enabled in Extensions.

License

MIT. See LICENSE.